This study was approved by the ethical committee of the Institutional Review Board of Kyung Hee University Hospital at Gangdong (No. 2015-02-006-001). Written informed consent was obtained by all participants enrolled. Twenty subjects were studied using two different sequences. Ten subjects (Group I) were scanned with a three-dimensional (3D)-segmented gradient-echo echo-planar imaging (EPI) sequence to obtain full Z-spectrum signals using a 3T MRI system (Achieva 3.0T; Philips, Amsterdam, Netherlands). All the subjects were woman. Ten other subjects (Group II) were scanned with a 3D gradient and spin-echo (GRASE) sequence to obtain full Z-spectrum signals using another 3T MRI system (Ingenia 3.0T; Philips). In this group, eight were women and two were men. The subjects were recruited for optimizing the CEST MRI techniques employed in Kyung Hee University Hospital at Gangdong to ultimately apply this technique for imaging the brain of patients with Alzheimer's disease. Table 1 lists the characteristics of the subjects and the CEST signal acquisitions.

The first full Z-spectrum data were acquired with the 3D-segmented gradient-echo EPI sequence13) using an eight-channel sensitivity-encoding (SENSE) coil. To induce CEST saturation exchange, we set the B1 amplitude of the saturation pulse as 1 µT; the saturation pulse duration per pulse as 70 ms; and the total number of shots to the center of the k-space as 126. Therefore, the total saturation length was 8.8 seconds. We obtained the full Z-spectrum via 29 dynamics at offset frequencies ranging from –6.00 ppm to 6.00 ppm with a continuously increasing frequency interval of 0.43 ppm. The first acquired image was the reference image S₀ obtained at an offset frequency of –40. The imaging parameters were as follows: repetition time (TR)/echo time (TE)=150/7.1 ms; acquisition matrix=112×112; acquisition voxel size=2×2×6.40 mm³; reconstruction voxel size=0.76×0.76×3.20 mm³; EPI factor=9; flip angle (FA)=7°; field of view (FOV)=220×220×06 mm³; SENSE factor=2 for the anterior-posterior direction and 1 for the right-left direction; number of slices=33; and imaging orientation=transverse. The scan time was 567 seconds.

The second full Z-spectrum data were acquired with the 3D GRASE sequence14) by using a 32-channel SENSE coil. To induce CEST saturation exchange, we set the B1 amplitude as 2 µT; the saturation pulse duration as 200 ms with a 10-ms interval between the pulses; and the number of saturation pulses as 4. Therefore, the total saturation length was 0.83 seconds. We obtained the full Z-spectrum via 37 dynamics at offset frequencies ranging from –5.00 ppm to 5.00 ppm with an alternatively increased frequency interval of 0.25 ppm at offset frequencies ranging from ±0.25 ppm to ±4.00 ppm and thereafter acquired the images at offset frequencies of ±4.5 ppm and ±5.0 ppm. The first acquired image was the reference image S₀ at –40 ppm, and the second acquired image was taken at an offset of 0 ppm with respect to the direct saturation of water. The imaging parameters were as follows: TR/TE=2200/16 ms; acquisition matrix=104×92; acquisition voxel size=2×2×8 mm³; reconstruction matrix size=1×1×4 mm³; FA=90°; SENSE factor=2 for the anterior-posterior direction and 1 for the right-left direction; turbo spin-echo (TSE) factor=23; EPI factor=7; number of slices=23; and imaging orientation=transverse. The scan time was 509 seconds.

Finally, for image registration and brain-tissue segmentation, sagittal structural 3D T1-weighted (T1W) images were acquired with the magnetization-prepared rapid acquisition of the gradient echo sequence with the following parameters: TR=8.1 ms; TE=3.7 ms; FA=8°; FOV=236×236 mm²; and voxel size=1×1×1 mm³. In addition, T2-weighted TSE and fluid-attenuated inversion recovery images were acquired to examine any brain malformations.

The MRI data were analyzed using MATLAB (http://www.mathworks.com) (MathWorks, Natick, MA, USA) and Statistical Parametric Mapping Version 12 (SPM12) (http://www.fil.ion.ucl.ac.uk/spm/) (Wellcome Trust Centre for Neuroimaging, London, UK). Fig. 1 shows the pre-processing steps of the 3D T1W and full Z-spectrum images. First, the 3D T1W and full Z-spectrum images were co-registered. Second, the 3D T1W images were segmented into gray and white matter using the CAT12 toolbox (http://www.neuro.uni-jena.de/cat/) (Structural Brain Mapping Group, Jena, Germany) to obtain brain-tissue compartments. Third, the co-registered full Z-spectrum images for each subject were spatially normalized into a Montreal Neurological Institute (MNI) brain template using the subject's deformation field information obtained from the brain-tissue-segmentation step of the 3D T1W image. Finally, the CSF signal in the full Z-spectrum images was masked using information from the segmented gray- and white-matter tissue compartments to evaluate the full Z-spectrum images while minimizing the contributions of the CSF signals. We only included voxels that had more than 50% gray and white matter because voxels with more than 50% CSF are usually of no interest in some clinical evaluations. After this preprocessing step, we obtained the MTR_asym map, which is the CEST asymmetry map.

We created two sets of MTR_asym maps for each offset frequency for each subject using the pre-processed full Z-spectrum data in the masked and unmasked conditions by employing the following steps. The spatially normalized full Z-spectrum images in the masked and unmasked conditions were divided by the reference image S₀ obtained at an offset frequency of –40 ppm. To determine the offset frequency for the minimum signal, the B₀ inhomogeneity correction step was performed using the 10th degree polynomial fitting,15) assuming that the actual water resonance was at the frequency with the lowest signal intensity and excluding the fat-dominate voxels. The water resonance frequency was estimated as the frequency with the lowest signal intensity from the fitted curve and shifted along the direction of the offset axis to 0 ppm at its lowest intensity on a voxel-by-voxel basis. The MTR_asym maps with respect to the water frequency under the masked and unmasked conditions were obtained using the following equation:316) Here, Δω is the frequency difference with respect to water, S_sat is the signal with the saturation pulse, and S₀ is the signal observed without any saturation pulse. Usually, the CEST effect from the solute protons is detected at a frequency lower than 6 ppm with respect to water. We used the MTR_asym maps obtained at 0.86, 2.14, 3.00, and 3.43 ppm for the 3D-segmented gradient-echo EPI data and those obtained at 1, 2, 3.0, and 3.5 ppm for the 3D GRASE data.

The MTR_asym maps obtained in the masked and unmasked conditions for each subject were smoothed using a Gaussian smoothing kernel at a full-width half maximum of 8×8×8 mm³. We compared the MTR_asym maps obtained in the masked and unmasked conditions using voxel-based and region-of-interest (ROI)-based methods. For the voxel-based analysis, the MTR_asym value in each voxel was compared between the masked and unmasked conditions using paired t-tests for each offset frequency and for each sequence, separately. A significance level of P=0.001 was applied without correcting for multiple comparisons in clusters with at least 30 contiguous voxels. This voxel-wise analysis was performed to define the ROIs in the brain.

The ROIs were defined at Brodmann area (BA) 10, BA21, cerebellar tonsil, declive, precuneus, and thalamus based on the results of the voxel-based analysis using the WFU PickAtlas toolbox (http://fmri.wfubmc.edu/software/pickatlas) (Radiology Informatics and Imaging Laboratory, Winston-Salem, NC, USA). BA10 is the anterior portion of the prefrontal cortex in the human brain, which includes the frontopolar, rostrolateral, and anterior prefrontal cortexes. BA21 is the part of the middle temporal gyrus. The ventricles were not selected as ROIs because many of the CSF signals in the ventricles were masked; hence, the differences in the ventricles were natural. The MTR_asym values for each ROI and at the four offset frequencies were obtained using the MarsBaR toolbox (http://marsbar.sourceforge.net/). The normal distribution of the MTR_asym values for each ROI was tested using the Kolmogorov-Smirnov method. Paired t-tests were used to compare the MTR_asym values between the masked and unmasked conditions for each ROI and each offset frequency. The results were regarded as significant if the two-tailed probability of the results was lower than 0.05. MedCalc (MedCalc Software, Ostend, Belgium) was used for the statistical analysis.

In the 3D-segmented gradient-echo EPI sequence, the MTR_asym values in the unmasked condition were higher than those in the masked condition at offset frequencies of 0.86 ppm and 2.14 ppm (Supplementary Fig. 1). However, the MTR_asym values in the unmasked condition were higher or lower than those in the masked condition at 3 ppm and 3.43 ppm frequencies as listed in Supplementary Table 1, which summarizes the result of the voxel-based comparisons of the MTR_asym maps between the masked and unmasked conditions in the 3D-segmented gradient-echo EPI sequence.

For the 3D GRASE sequence, only a few areas showed significant differences in the MTR_asym values between the two conditions. The MTR_asym values were greater in the unmasked condition than in the masked condition for the right inferior semi-lunar lobule (Talairach coordinates=1.13, –58.93, and –36.42; cluster size=66; Z-score=3.29) at an offset frequency of 2.14 ppm offset frequency and for the left precuneus (Talairach coordinates= 0.32, –64.94, and 25.16; cluster size=434; Z-score= 3.89) at an offset frequency of 3.00 ppm. No significant differences were observed in the MTR_asym values between the two conditions at other offset frequencies. Finally, all areas showed lower MTR_asym values in the masked condition than in the unmasked condition.

Table 2 summarizes the results of the ROI-based comparisons of the MTR_asym values between the masked and unmasked conditions for each ROI and for each offset frequency obtained with the 3D-segmented gradient-echo EPI sequence. The MTR_asym values were higher in the unmasked condition than in the masked condition at offset frequencies of 0.86, 2.14, 3.00, and 3.43 ppm. The MTR_asym values were significantly different between the masked and unmasked conditions for the following brain regions: BA21, tonsil, and declive at 0.86 ppm; tonsil and thalamus at 2.14 ppm, indicating the presence of guanidium protons; BA10 and BA21 at 3.00 ppm, indicating the presence of amines; and BA10, BA21, and precuneus at 3.43 ppm.

Table 3 summarizes the results of the ROI-based comparisons of the MTR_asym values between the masked and unmasked conditions for each ROI and for each offset frequency obtained with the 3D GRASE sequence. Significant differences in the MTR_asym values between the two conditions were observed for the following brain areas: all ROIs except the precuneus at 1.0 ppm, indicating the presence of hydroxyl protons; all ROIs except BA10 and thalamus at 2.0 ppm; all ROIs except thalamus at 3.0 ppm; and all ROIs except BA21 and thalamus at 3.5 ppm, indicating the presence of amides.

We found significant contributions of the CSF signals to the MTR_asym map, which overestimated the MTR_asym values in the elderly human brain. The MTR_asym values were altered with the masking of the CSF signals, irrespective of the saturation offset frequency with respect to water. In elderly subjects, the overestimation of the MTR_asym values due to the changes in the CSF signals could be attributed to brain atrophy, which results in widening of the perivascular spaces, which are basically CSF spaces. To obtain accurate CEST effects in the elderly brain, the CSF signal should be masked either using some post-processing technique or using some data acquisition method, e.g., an inversion-recovery preparation. This corrected MTR_asym value might be valuable for evaluating brain metabolites in patients with neurodegenerative diseases.1718) For the amide and amine protons, the results of the ROI-based analyses showed that the MTR_asym values for the thalamus were not significantly different between the two conditions for both the sequences, whereas the values were significantly different for the precuneus. Several previous studies have reported brain atrophy at the precuneus in elderly subjects, but not at the thalamus.11) Our result, therefore, showed that the difference in the MTR_asym values between the two conditions could be related to CSF contributions in the CEST signals rather than to an artifact. Some regions showed a difference in the MTR_asym values between the unmasked and masked conditions only at the hydroxyl offset frequency. This indicates that it is difficult to accurately measure the MTR_asym value at this offset frequency using a 3T MRI system because of the proximity to water.

The MTR_asym values acquired using the two sequences were sensitive to the contribution of the CSF signals. For the 3D-segmented gradient-echo EPI sequence, we used a relatively lower power saturation pulse than that used for the 3D GRASE sequence. At the amide offset frequency, the MTR_asym values were significantly different between the two conditions for BA10, BA21, and precuneus in the case of the EPI sequence, but for BA10, tonsil, declive, and precuneus in the case of the GRASE sequence. At the amine offset frequency, the MTR_asym values were significantly different between the two conditions only for BA10 and BA21 with the EPI sequence, but at all the ROIs except the thalamus with the GRASE sequence. These differences could be attributed to the fact that the 3D GRASE sequence more strongly reflects the effect of the CSF signal than the 3D EPI sequence. The GRASE sequence is used in multiple spin-echo pulses. Another probable reason more ROIs showed significant differences between the masked and unmasked conditions in the GRASE sequence than they did in the segmented EPI sequence would be the quality of the full Z-spectrum. The full Z-spectrum with the 3D GRASE sequence was much broader than that with the 3D EPI sequence. It may be necessary to evaluate the MTR_asym value using a relatively low B1 power of the saturation pulse for the 3D GRASE sequence, which may slightly increase the sharpness of the full Z-spectrum; however, a previous study showed that the full Z-spectrum with a B₁ amplitude of 1 µT was still broad.19) When using a low B₁ power, the effect of the contribution of CSF signals to the MTR_asym value may therefore be similar to that in this study. Nevertheless, further studies with different types of pulse sequences should be performed to validate these findings.

First, we only evaluated the full Z-spectrum data acquired using a pulsed-saturation RF preparation.2021) Because saturation with a continuous-wave RF pulse is more effective than that with the pulse-wave RF, the effect of CSF signals on the full Z-spectrum data needs to be evaluated. However, a continuous-wave RF pulse for saturation cannot be generated on most clinical scanners owing to hardware limitations. Second, in this study, the B₁ amplitude was not optimized for amide protons or amine because we did not target one specific proton. The saturation preparation depends only on the B1 amplitude and saturation duration. A low power is recommended for the saturation of amide protons, but a relatively high power is recommended for the saturation of amine protons. For the 3D GRASE sequence, the B1 power for the saturation transfer may have been too high because the full Z-spectrum obtained was broad. Third, the B₀ correction method using the 10th degree polynomial fitting may not be optimal. Each voxel has a small difference in precession frequency due to structural susceptibility, which is particularly significant at high magnetic fields.5) B₀ inhomogeneity leads to a shift in the water resonance frequency, resulting in direct asymmetric water saturation effects, and thus, artificial CEST effects in an asymmetry analysis. A better method may have to be applied for correcting the B₀ inhomogeneity.415) Fourth, we used the MTR_asym maps obtained at 0.86, 2.14, 3.00, and 3.43 ppm for the 3D-segmented gradient-echo EPI data and those obtained at 1, 2, 3.0, and 3.5 ppm for the 3D GRASE data. The slight differences in the offset frequencies in the two sequences may not represent the same exchangeable protons for both measurements. In this study, we did not focus on evaluating the interpretation method after field inhomogeneity correction. Finally, we included a relatively small population in this study. Further large-scale studies should be conducted to validate our observations.